The ad account reports a third more conversions than the warehouse, every single month

Four things produce a persistent gap between an ad account and a warehouse, and they are worth ranking before anyone opens a query editor. The 2 numbers are dated differently. Part of the platform's figure is modelled rather than counted. The platform is counting a single conversion twice. The 2 sides are measuring windows of different length. Deduplication and window length are engineering defects with fixes. Date basis and modelling are definitional differences between 2 systems that were never built to agree, and the correct response to them is a documented tolerance.

Say the last part out loud early, because the expensive version of this problem is not the gap. It is the meeting: a monthly hour in which a performance team and a data team re-derive the same 4 explanations from scratch, and nobody writes anything down. An exact match is not an available outcome. A stable, explained, bounded difference is.

Confirm the shape of the gap before anyone explains it

Month totals cannot be diagnosed, because every cause is present in them at once. Take one day and one campaign, and cut both sides the same way. An afternoon spent here removes at least 2 of the 4 candidates.

1. Pick a single completed day at least 10 days in the past, so the platform's own figures have finished settling, and a single campaign with enough volume to be readable — 50 conversions is a workable floor.
2. Pull the platform's number twice: once dated by conversion date, once dated by the date of the click that got credit. Most interfaces offer both. The difference between those 2 pulls is the size of your date-basis problem, measured rather than argued.
3. Pull the warehouse number for the same day from the raw export, not from a dashboard, and confirm the day's partition is complete — provisional data is its own class of confusion, covered in yesterday's numbers moving after the table lands.
4. Bucket the warehouse rows 3 ways: by transport (browser only, server only, both), by consent state at the time of the event, and by whether a click identifier was present. Three counts, one query.
5. Compare the both-transports count against the platform's figure. If the platform number exceeds the sum of your distinct conversions, you have duplication. If it exceeds your rows carrying a click identifier but not your total, you are looking at modelling and matching loss.
6. Write the 5 numbers into a table and keep it. The next time this argument starts, it starts from evidence.

The four causes, ranked

Rank matters because the investigation order is the opposite of the intuitive one. A stable gap that has been the same size for 6 months is almost never duplication — duplication arrives on a date. A gap that changed shape last Tuesday is almost never date basis, because date basis does not change unless someone changed a report.

The two you can close: duplication and window length

Duplication is the only cause that makes the platform figure wrong rather than different. It appears as a step change on the day server-side sending was enabled, with order volume flat underneath it, and it is isolated to one of a small number of failure points in conversions counting twice after the server goes live. Rule it out before touching anything else, because every other explanation is contaminated while it is present.

Window length is duller and more common. The platform is configured with a click-to-conversion lookback; your warehouse query joins conversions to clicks with whatever window someone typed 2 years ago, and the 2 have never been compared. Read the configured value from the account, put it in the model as a named parameter rather than a literal, and note it in the report header. If the model itself is ad hoc, this is a good moment to build it properly — the layering is in turning raw event rows into a model people can query.

The two you can only account for: date basis and modelling

Date basis is the largest cause of a stable gap and the least interesting once understood. Ad platforms credit a conversion to the day of the click that earned it, because that is the day the money was spent. A warehouse records the day the order was placed. For a business with a same-day purchase cycle these nearly agree; for one with a 3-week consideration window they never will, and the gap is largest in the most recent 2 weeks — exactly the period people look at. Restating the warehouse query on click date is legitimate and takes an hour. What is not legitimate is comparing a click-dated number to a conversion-dated one and calling the difference a data quality problem.

Modelling is the harder conversation. Where a conversion cannot be observed — no identifier, no consent for storage, a browser that dropped the join — platforms estimate how many occurred and report the estimate alongside the observed count. No row exists underneath that figure, which is why it cannot be exported, joined or matched; the mechanism is set out in why modelled conversions never reconcile to rows. The practical consequence is a floor: your gap will not close below the modelled share, however good your pipeline gets.

Two systems that were built to answer different questions will not agree by being asked harder. The deliverable is not a match. It is a difference small enough, stable enough and explained well enough that nobody has to re-derive it next month.

What a tolerable gap looks like, and how to write it down

There is no industry constant here, and anyone quoting one is quoting a number from a different business. What is transferable is the shape of the judgement. We treat stability as the signal rather than size: a gap that sits within a 5-point band for 3 consecutive months is a definitional difference, and one that swings 15 points between months is a defect regardless of its average. Size only matters relative to how much of the platform figure is modelled and how long the consideration window is.

• State the date basis on every report, in the header, in words. Most of these arguments are 2 people comparing correctly-computed numbers that were never comparable.
• Publish the observed and modelled split where the platform exposes it, so the irreducible part of the gap is visible rather than argued about.
• Track the gap as a monitored series with an alert on the change, not the level. A jump is a defect; a level is a definition.
• Automate the monthly reconciliation. It is a scheduled query, a variance check and a written explanation — bounded, repeatable work of the kind we describe in AI agents and automation, with a person still reading the output.
• Recheck the tolerance whenever an attribution setting changes, since the band is only valid for the configuration it was measured under — the discipline in changing attribution models mid-year.

Where the tree ends: fix, restate, or document

1. Fix. The platform figure exceeds your distinct order count, or the gap appeared on a datable event. This is duplication or a misconfigured window, and it is a pipeline change with a testable outcome.
2. Restate. The gap tracks the difference between click date and conversion date, and shrinks as older days fill in. Re-cut the warehouse query on click date, publish both, and stop comparing the wrong pair.
3. Document. What remains after the first 2 is modelling plus matching loss. Write the tolerance into the report as a stated band with the date it was measured, and review it quarterly rather than monthly.

The seller side of advertising has the same structural lesson, in a different costume: when a buyer's records disagree with yours about what you are entitled to sell, the fix is not counting harder but correcting a declared record both parties read, which is what sits behind buyers flagging our inventory as unauthorised. Reconciliation work is nearly always about agreeing a definition, then encoding it, then leaving it alone.

If this work is going outside — and it often should, because it is unglamorous and specific — the questions worth asking are the ones in choosing a development partner. The rest of the decisions around the pipeline sit under tracking, consent and event pipelines, part of our marketing and advertising practice.

• reconciliation
• attribution
• conversion tracking
• reporting